RFR: 8353013: java.net.URI.create(String) may have low performance to scan the host/domain name from URI string when the hostname starts with number

Mikhail Yankelevich myankelevich at openjdk.org
Fri Apr 11 18:41:39 UTC 2025


On Fri, 28 Mar 2025 15:19:46 GMT, Rohitash <duke at openjdk.org> wrote:

> `scanByte` throws `NumberFormatException` for URIs that start with numbers, e.g., https://11111111.x.y/
> The current flow is `parseIPv4Address` → `scanIPv4Address` → `scanByte`. `parseIPv4Address` uses `NumberFormatException` for control flow, so it captures the exception, ignores it, and returns -1. This has been reported by AWS customer to cause low performance. Details: [JDK-8353013](https://bugs.openjdk.org/browse/JDK-8353013) & https://github.com/aws/aws-sdk-java-v2/issues/5933
> 
> This PR avoids NumberFormatException by skipping calls to `Integer.parseInt` if the number of digits in the octet is > 3.
> 
> I benchmarked on local machine for potential regressions.
> https://gist.github.com/rk-kmr/cb1a3d59225c17b180a29cc125ebf887
> 
> 
> Benchmark                                     Mode     Cnt        Score       Error   Units
> URIBenchMark.newImplWithNormalUrl            thrpt      10        0.004 ±     0.001  ops/ns
> URIBenchMark.newImplWithNumberlUrl           thrpt      10        0.004 ±     0.001  ops/ns
> URIBenchMark.oldImplWithNormalUrl            thrpt      10        0.004 ±     0.001  ops/ns
> URIBenchMark.oldImplWithNumUrl               thrpt      10        0.001 ±     0.001  ops/ns
> URIBenchMark.newImplWithNormalUrl             avgt      10      236.762 ±     8.700   ns/op
> URIBenchMark.newImplWithNumberlUrl            avgt      10      264.017 ±     7.274   ns/op
> URIBenchMark.oldImplWithNormalUrl             avgt      10      233.853 ±     6.539   ns/op
> URIBenchMark.oldImplWithNumUrl                avgt      10     1183.572 ±    29.242   ns/op
> 
> 
> I ran following tests.
> 
> make test-tier1
> make test-tier2
> make test TEST=jdk/java/net

Could you please also add  a test for this change/ add a bug id to the existing test if there already is one covering this functionality. Just in case

src/java.base/share/classes/java/net/URI.java line 3429:

> 3427:             int q = scan(p, n, L_DIGIT, H_DIGIT);
> 3428:             if (q <= p) return q;
> 3429: 

Nit: Copyright year

src/java.base/share/classes/java/net/URI.java line 3435:

> 3433: 
> 3434:             // Calculate the number of significant digits (after leading zeros)
> 3435:             int significantDigits = q - i;

Nit: Do you think it might be more clear if this is named something like `significantDigitsNum` or similar? No strong feelings, just confused me a little when I was reading this.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/24295#issuecomment-2766357418
PR Review Comment: https://git.openjdk.org/jdk/pull/24295#discussion_r2020971319
PR Review Comment: https://git.openjdk.org/jdk/pull/24295#discussion_r2020969734


More information about the net-dev mailing list